aacenc: use av_clip() instead of av_clip() during quantization
authorRostislav Pehlivanov <atomnuker@gmail.com>
Tue, 29 Mar 2016 23:58:28 +0000 (00:58 +0100)
committerRostislav Pehlivanov <atomnuker@gmail.com>
Wed, 30 Mar 2016 00:04:43 +0000 (01:04 +0100)
Seems like clang might be miscompiling it and causing a signed integer overflow,
making a FATE test fail.
Doesn't seem to affect performance, it only runs on the ESC codebook.

Reviewed-by: Claudio Freire <klaussfreire@gmail.com>
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
libavcodec/aacenc_quantization.h

index 4250407..106e632 100644 (file)
@@ -141,7 +141,7 @@ static av_always_inline float quantize_and_encode_band_cost_template(
             if (BT_ESC) {
                 for (j = 0; j < 2; j++) {
                     if (ff_aac_codebook_vectors[cb-1][curidx*2+j] == 64.0f) {
-                        int coef = av_clip_uintp2(quant(fabsf(in[i+j]), Q, ROUNDING), 13);
+                        int coef = av_clip(quant(fabsf(in[i+j]), Q, ROUNDING), 0, (1 << 13) - 1);
                         int len = av_log2(coef);
 
                         put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);